[Docs][Serve] Speed up weights loading by AMI and Docker Image

[cblmemo]

TODO: Benchmark and get some numbers
Fix bug using

sudo jq '.["exec-opts"] = ["native.cgroupdriver=cgroupfs"]' /etc/docker/daemon.json > /tmp/daemon.json && sudo mv /tmp/daemon.json /etc/docker/daemon.json
sudo systemctl restart docker

Tested (run the relevant ones):

• Code formatting: bash format.sh
• Any manual or new tests for this PR (please specify below)
  • sky launch the following yaml and works properly

# mixtral.yaml
service:
  readiness_probe:
    path: /v1/chat/completions
    post_data:
      model: mistralai/Mixtral-8x7B-Instruct-v0.1
      messages:
        - role: user
          content: Hello! What is your name?
      max_tokens: 1
    initial_delay_seconds: 1800
  replicas: 1

# Fields below describe each replica.
resources:
  cloud: gcp
  image_id: docker:cblmemo/mixtral-vllm:latest
  ports: 8080
  accelerators: {L4:8, A10g:8, A100:4, A100:8, A100-80GB:2, A100-80GB:4, A100-80GB:8}

setup: |
  conda activate vllm
  if [ $? -ne 0 ]; then
    conda create -n vllm python=3.10 -y
    conda activate vllm
  fi
  pip install vllm==0.3.0 transformers==4.37.2

run: |
  conda activate vllm
  export PATH=$PATH:/sbin
  python -m vllm.entrypoints.openai.api_server \
    --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \
    --host 0.0.0.0 --port 8080 \
    --model mistralai/Mixtral-8x7B-Instruct-v0.1

• All smoke tests: pytest tests/test_smoke.py
• Relevant individual smoke tests: pytest tests/test_smoke.py::test_fill_in_the_name
• Backward compatibility tests: bash tests/backward_comaptibility_tests.sh

[Michaelvll]

It seems the benefit of using machine image or docker image is more for reducing the setup time, instead of the the model weight downloading time, as they are mostly for packaging the dependencies and not neccessarily means it will speed up the download of the image or the model, which should be limited by the network bandwidth instead.

Should we rewrite the section as reducing the overhead of environment setup?

[cblmemo]

Good point! Let me rephrase that

[cblmemo]

I guess there are some speedups if using a machine image? That should be optimized by the cloud provider which makes it faster than a plain network download?

[Michaelvll]

Machine images are intended to be used in a single region. It can be used in another region for launching a VM, but there will involve data transfer from one region to another. Cloud provider should have optimized it, but we probably want to be careful about the wording to avoid having a impression that all the benefits comes from weight loading.

[concretevitamin]

Have we actually timed these two methods?

[cblmemo]

I'm feeling like it is hard to make a fair comparison - it is largely dependent on the base docker/machine image used... Though I'll try to make some benchmarks and see the results 🫡

[Michaelvll]

We could mention something about how this docker image should be built, especially, it could have SkyPilot runtime pre-built. Something like the following would be useful (could you help giving an concrete example for how to install vllm, and download the image in the Dockerfile below for a better reference, i.e. replacing the line # Your dependencies installation and model download code goes here with actual workable commands for serving vllm+mistral):

Your docker image can have all skypilot dependencies pre-installed to further reduce the setup time, you could try building your docker image based from our base image. The `Dockerfile` could look like the following:
```Dockerfile
FROM docker:berkeleyskypilot/skypilot-k8s-gpu:latest

# Your dependencies installation and model download code goes here
```

[Michaelvll]

Thanks for adding this doc @cblmemo! We have users asking for this and it would be nice we can directly point them to this page. : )

[cblmemo]

Thanks for adding this doc @cblmemo! We have users asking for this and it would be nice we can directly point them to this page. : )

Just want to update the status of this PR first: I found a mysterious bug that causes an NVML initialization error when using docker container as runtime env. By bisect it seems like those lines are causing the error:

skypilot/sky/provision/docker_utils.py

Lines 271 to 274 in 82c50f5

	'sudo systemctl stop jupyter > /dev/null 2>&1 || true;'
	'sudo systemctl disable jupyter > /dev/null 2>&1 || true;'
	'sudo systemctl stop jupyterhub > /dev/null 2>&1 || true;'
	'sudo systemctl disable jupyterhub > /dev/null 2>&1 || true;',

That is very strange since those are running on the host but somehow affect the containers. Will investigate more.

[Michaelvll]

We should consider having this PR updated and merged as well. : )

[cblmemo]

We should consider having this PR updated and merged as well. : )

This PR is blocked by the max/ultra disk tier as the current performance is not better than install everything from pip...

[Michaelvll]

Another user requests this. : )

[cblmemo]

Left some benchmark results using PR #3860. In the following table, high = gp3 7,000 IOPS, ultra = io2 20,000 IOPS, max = io2 100,000 IOPS. All tests are running on AWS and the result is the e2e execution time for launching a Llama 2 70b checkpoint w/ the latest version of vLLM, on an A10G:8 instance.

	high	ultra	max
Use AMI	> 2 hours	487s	467s
Download from HF	524s	410s	-

In conclusion, our high disk tier is indeed not enough for large checkpoint downloading and the ultra tier increases the performance a lot. Though the AMI does not enhance the performance as expected; there might be other bottlenecks like the download speed of the AMI.

[Michaelvll]

We should revamp this PR with our latest findings and support for ultra disk : )

[cblmemo]

We should revamp this PR with our latest findings and support for ultra disk : )

Done in #3949 . Still keeping this so we could investigate if it is possible to speed up by using AMI & docker image.